[llvm] d4b3380 - [ARM] Handle Splats in MVE lane interleaving

[David Green via llvm-commits]

Author: David Green
Date: 2021-03-30T11:19:16+01:00
New Revision: d4b3380dfe62be6e2f1994e821d4d777e84e4294

URL: https://github.com/llvm/llvm-project/commit/d4b3380dfe62be6e2f1994e821d4d777e84e4294
DIFF: https://github.com/llvm/llvm-project/commit/d4b3380dfe62be6e2f1994e821d4d777e84e4294.diff

LOG: [ARM] Handle Splats in MVE lane interleaving

As another addition to MVE lane interleaving, this handles Splat shuffle
vectors, as the shuffle of a splat is a splat.

Differential Revision: https://reviews.llvm.org/D97291

Added: 
    

Modified: 
    llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
    llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp b/llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
index ce01245fbd14..9189298a9d04 100644
--- a/llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
+++ b/llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
@@ -212,6 +212,12 @@ static bool tryInterleave(Instruction *Start,
         Worklist.push_back(cast<Instruction>(Use));
       break;
 
+    case Instruction::ShuffleVector:
+      // A shuffle of a splat is a splat.
+      if (cast<ShuffleVectorInst>(I)->isZeroEltSplat())
+        continue;
+      LLVM_FALLTHROUGH;
+
     default:
       LLVM_DEBUG(dbgs() << "  Unhandled instruction: " << *I << "\n");
       return false;

diff  --git a/llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll b/llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
index 883a0781f313..db9c286a44be 100644
--- a/llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
@@ -422,17 +422,15 @@ entry:
 define arm_aapcs_vfpcc void @mul_i16(<8 x i16> *%A, <8 x i16> *%B, i32 %C, <8 x i16> *%D) {
 ; CHECK-LABEL: mul_i16:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vldrh.s32 q0, [r1]
-; CHECK-NEXT:    vldrh.s32 q1, [r0]
-; CHECK-NEXT:    vldrh.s32 q2, [r0, #8]
+; CHECK-NEXT:    vldrw.u32 q0, [r1]
+; CHECK-NEXT:    vldrw.u32 q1, [r0]
 ; CHECK-NEXT:    rsbs r2, r2, #0
-; CHECK-NEXT:    vmul.i32 q0, q1, q0
-; CHECK-NEXT:    vldrh.s32 q1, [r1, #8]
+; CHECK-NEXT:    vmullt.s16 q2, q1, q0
+; CHECK-NEXT:    vmullb.s16 q0, q1, q0
+; CHECK-NEXT:    vshl.s32 q2, r2
 ; CHECK-NEXT:    vshl.s32 q0, r2
-; CHECK-NEXT:    vmul.i32 q1, q2, q1
-; CHECK-NEXT:    vstrh.32 q0, [r3]
-; CHECK-NEXT:    vshl.s32 q1, r2
-; CHECK-NEXT:    vstrh.32 q1, [r3, #8]
+; CHECK-NEXT:    vmovnt.i32 q0, q2
+; CHECK-NEXT:    vstrw.32 q0, [r3]
 ; CHECK-NEXT:    bx lr
 entry:
   %a = load <8 x i16>, <8 x i16> *%A, align 4
@@ -451,17 +449,15 @@ entry:
 define arm_aapcs_vfpcc void @mul_i8(<16 x i8> *%A, <16 x i8> *%B, i16 %C, <16 x i8> *%D) {
 ; CHECK-LABEL: mul_i8:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vldrb.s16 q0, [r1]
-; CHECK-NEXT:    vldrb.s16 q1, [r0]
-; CHECK-NEXT:    vldrb.s16 q2, [r0, #8]
+; CHECK-NEXT:    vldrw.u32 q0, [r1]
+; CHECK-NEXT:    vldrw.u32 q1, [r0]
 ; CHECK-NEXT:    rsbs r2, r2, #0
-; CHECK-NEXT:    vmul.i16 q0, q1, q0
-; CHECK-NEXT:    vldrb.s16 q1, [r1, #8]
+; CHECK-NEXT:    vmullt.s8 q2, q1, q0
+; CHECK-NEXT:    vmullb.s8 q0, q1, q0
+; CHECK-NEXT:    vshl.s16 q2, r2
 ; CHECK-NEXT:    vshl.s16 q0, r2
-; CHECK-NEXT:    vmul.i16 q1, q2, q1
-; CHECK-NEXT:    vstrb.16 q0, [r3]
-; CHECK-NEXT:    vshl.s16 q1, r2
-; CHECK-NEXT:    vstrb.16 q1, [r3, #8]
+; CHECK-NEXT:    vmovnt.i16 q0, q2
+; CHECK-NEXT:    vstrw.32 q0, [r3]
 ; CHECK-NEXT:    bx lr
 entry:
   %a = load <16 x i8>, <16 x i8> *%A, align 4